Research On Parameters And Structural Optimization Of Convolutional Neural Networks

Convolutional neural network is a kind of deep neural network which is composed of convolutional computation and has excellent ability of image feature extraction.Since the performance difference of the network is mainly reflected in training of parameter and design of structure of the network,optimization of parameters and structure can effectively improve the Convolutional Neural Network.This paper mainly studies two aspects of parameter and structure optimization of convolutional neural network.In parameters,this paper proposes a parameterized optimization method based on image features.This method uses multiple groups of image features to initialize the convolution kernel,optimizes the parameters and accelerates the convergence of the network by adjusting the initial kernel of the convolution.In structure,this paper proposes a convolutional neural network based on mean iteration threshold segmentation method which is determined by prior knowledge in data,so the network has fewer redundant parameters and less dependence for data.The specific work is as follows:1.In order to solve the problem that the current convolution kernel initialization method is easy to cause the slow convergence speed,which based on image features is proposed in this paper This method sample data with combines fuzzy and edge technology to extract the feature block with the principal component analysis,which finish the convolution kernel.It not only solves the limitation of the principal component analysis on network structure,but also enhances to extract feature.This method is applied to Cifar-10 and Corel-1000 data sets to compare with Gaussian method and He method in paper which show the experimental effect of this method is better than other convolution kernel initialization methods.2.A convolutional neural network based on mean iteration threshold segmentation method is proposed to solve the problem of large data dependence of convolutional neural network.In this paper,the mean iteration threshold segmentation method is used to filter the image background,and then the network structure is designed based on Alex Net to reduce the redundant network parameters and reduce the degree of dependence of network for data.In paper,this network is applied to small-scale data sets to compare it with Alex Net and VGGNet,which show this network has ideal recognition effect in small tasks and better than other universal convolutional neural networks.